Future changes in flash flood frequency and magnitude over the European Alps

Zander, Marjanne J.; Viguurs, Pety J.; Sperna Weiland, Frederiek C.; Weerts, Albrecht H.

doi:https://doi.org/10.5194/hess-2023-274

Preprints

https://doi.org/10.5194/hess-2023-274

Preprints

05 Dec 2023

| 05 Dec 2023

Status: this discussion paper is a preprint. It has been under review for the journal Hydrology and Earth System Sciences (HESS). The manuscript was not accepted for further review after discussion.

Future changes in flash flood frequency and magnitude over the European Alps

Marjanne J. Zander, Pety J. Viguurs, Frederiek C. Sperna Weiland, and Albrecht H. Weerts

Abstract. Flash Floods are damaging natural hazards which often occur in the European Alps. Precipitation patterns and intensity may change in a future climate affecting their occurrence and magnitude. For impact studies, flash floods can be difficult to simulate due the complex orography and limited extent duration of the heavy rainfall events which trigger them. The new generation convection-permitting regional climate models (CP-RCMs) improve the representation of the intensity and frequency of heavy precipitation. Within CP-RCMs deep convection is resolved rather than parameterized. Therefore, this study combines such simulations with high-resolution distributed hydrological modelling to assess changes in flash flood frequency over the Alpine domain. We use output from a state-of-the-art CP-RCM to drive a high-resolution distributed hydrological wflow_sbm model covering most of the Alpine mountain range on an hourly resolution. First, the hydrological model was validated by comparing ERA5 driven simulation with streamflow observations from 130 stations (across Rhone, Rhine, Po, Adige and Danube basins). Second, a hourly wflow_sbm simulation driven by a CP-RCM downscaled ERAInterim simulation was compared to databases of past flood events to evaluate if the model can accurately simulate flash floods and to determine a suitable threshold definition for flash flooding. Finally, simulations of the future climate RCP 8.5 for the end-of-century (2096–2105) and current climate (1998–2007) are compared for which the CP-RCM is driven by a Global Climate Model. The simulations are compared to assess if there are changes in flash flood frequency and magnitude using a threshold approach. Results show a similar flash flood frequency for autumn in the future, but a decrease in summer. However, the future climate simulations indicate an increase in the flash flood severity in both summer and autumn leading to more severe flash flood impacts.

Received: 15 Nov 2023 – Discussion started: 05 Dec 2023

Download & links

Marjanne J. Zander, Pety J. Viguurs, Frederiek C. Sperna Weiland, and Albrecht H. Weerts

Status: closed

RC1:
'Comment on hess-2023-274', Anonymous Referee #1, 22 Dec 2023
This study investigates the potential changes in flash floods over the Alps due to climate change. The aim of this study was to evaluate the use of a convection-permitting regional climate model in combination with a distributed hydrological model to assess future changes in the frequency and magnitude of flash floods over the Alps. The UM convection-permitting model was used to project future climate at high spatial and temporal resolution. These projections were used as inputs to the distributed hydrological model wflow_sbm. The ability of the hydrological model to simulate historical flash floods was first assessed using reanalysis data. The modelling framework was then used to investigate future changes in flash floods.
Although the topic of the paper is very interesting and has not yet been addressed in the literature, I do not believe that the methodology is adequate to address the research question formulated in the introduction. The modelling framework and data set used in this study are not robust enough to support the conclusions.
Major comments:
The ability of the hydrological model to reproduce historical flash floods in the study area is very low. Apart from the fact that the methodology for assessing this is questionable (see comments below), the results show that the model is not able to detect floods based on the threshold approach for about half of the stations studied (Fig. 5). The hydrographs in Figure 3 also show the limited performance in simulating most floods (and not just flash floods). If the modelling framework does not capture the dominant processes triggering floods in this region in the historical period, it is very unlikely that it will do so under changing conditions. There could be several reasons for this (e.g. uncertainty in the input data). It could also be argued that a better modelling framework has not yet been developed. However, since the current modelling capabilities (at least for the modelling framework of this study) are not robust and reliable enough in this context, I do not think it should be used to study future changes in flash floods.

The definition of a flash flood lacks a very important feature: the temporal dynamics of the storm and/or flood event. To determine what is a flash flood, the authors use the specific peak discharge and the size of the upstream catchment. This definition was adapted from Amponsah et al. (2018) by removing the storm duration as a selection criterion (not mentioned in the manuscript). This means that any discharge exceeding the specified threshold and occurring in a catchment of less than 3000 km² is considered a flash flood, regardless of whether it was triggered by a uniform precipitation event lasting a few days or by a very intense and highly spatially variable convective storm event. Floods associated with slow catchment dynamics can theoretically be considered as flash floods in this modelling framework.

The validation/evaluation methodology is not suitable for assessing the ability of the modelling framework to project future changes in flash floods.
The aggregation to the daily time step for model validation does not allow to evaluate the ability of the model to simulate flash floods, because the hydrological processes involved affect the flow at sub-daily time steps.

Basing the performance assessment on peak flows and the KGE calculated over the entire time series is not sufficient to assess whether the modelling framework was able to capture the dominant flood generation processes.
As the catchments do not necessarily only experience flash floods, the number of points related to these floods used to calculate an error criterion is small compared to the rest of the time series. Therefore, a high KGE value in this case does not mean that the model performs well in simulating flash floods. In addition, the KGE values are quite low for many stations (KGE< 0.6 for about ½ of the stations, see Table 3, Crochemore et al. 2015).

Peak flow analyses rely on single points, which can be highly uncertain.

As part of the validation/evaluation analysis is based on nine historical floods, it would have been possible to plot the flood hydrographs and compare the simulations with the observations (if available).

The paper lacks a comprehensive description of the streamflow dataset, in particular why this dataset is suitable for answering the research question. It is not clear why only 130 gauged stations were used in this study and how they were selected.
Were the stations selected because their catchments experience flash floods? Were they chosen because of data availability, low human influence...?

What are the hydrological processes affecting river flows in these catchments?

Are all the catchments and sub-catchments prone to flash floods and how often compared to the other floods?

Why were stations with different temporal resolutions chosen?

A large part of the study area is not covered by gauged stations.

No bias correction was applied to the climate projections. This point is discussed in Section 4, the main arguments being that bias correction can distort the change signal and that the observational data available over the study region do not have the resolution of the CPM. However, most hydrological impact studies apply some form of bias correction, as biases are usually quite large at the hydrological scale (e.g. Teutschbein and Seibert, 2012). It is not clear how the choice not to apply a bias correction affects the results of this study, as there is no assessment of the climate variables for the historical period compared to the observations. Are the biases large enough to significantly change peak discharge simulations?

An analysis of changes in flood drivers might have been expected. Other studies reporting on the potential drying of the region are mentioned (l276 to 280), but the states and fluxes of the hydrological model could have been used for a more in-depth analysis.

As I understand it, one of the objectives of this study was to assess the potential added value of using a CPM in combination with a distributed hydrological model to assess future changes in flash floods over the Alps. In order to assess the added value of such a framework, this study should have included a benchmark, such as the projections of a regional climate model at 0.11° resolution.

Other comments
Some of the figures and tables do not highlight the results well enough (suggestions for improving the figures below).
2: add the delineation of the study region.

3: one hydrograph per line; extend in width; eventually show a smaller period.

5: increase resolution; change “f1” and “peirce_ss” to “F1” and “Peirce”. Add number of stations in the legend. Add number of stations with skill score = 0.

2: add a column with observed peak specific discharge.

6: apply a transformation on the values of the x-axis to improve visualisation.

L11-12: “and to determine a suitable threshold definition for flash flooding.” I did not understand why the definition of a threshold for flash floods is one of the objectives and why it is based on modelling results and not on observations.

L148-156: why is glacier modelling mentioned in these lines? The authors show that glaciers do not have a significant influence on the occurrence of flash floods, but do not conclude on the implications for their modelling framework.

L 198: the regional approach to flash flood validation should be better explained, e.g. with a simple figure showing how the threshold approach is used to determine whether the model detects a flash flood or not.

L205: “rare extreme events”. If the flash floods considered in this study are rare events, the limited duration of the CPM projections (10 years) may not allow such events to be studied. I would have expected a small analysis of the flood events that occurred in the historical period for the 130 stations.

L 268: “5.6m3s−1km−2 compared to 4.49m3s−1km−2”. The difference between the two values is most likely within the uncertainty range of the hydrological model (to be confirmed with the observed peak values to be added to Table 2).

L276 to 280 could be moved to the discussion section.

References:
Louise Crochemore, Charles Perrin, Vazken Andréassian, Uwe Ehret, Simon P. Seibert, Salvatore Grimaldi, Hoshin Gupta & Jean-Emmanuel Paturel (2015) Comparing expert judgement and numerical criteria for hydrograph evaluation, Hydrological Sciences Journal, 60:3, 402-423, DOI: 10.1080/02626667.2014.903331
Teutschbein, C., & Seibert, J. (2012) Bias correction of regional climate model simulations for hydrological climate-change impact studies: Review and evaluation of different methods. Journal of hydrology, 456, 12-29, DOI: 10.1016/j.jhydrol.2012.05.052
Citation: https://doi.org/10.5194/hess-2023-274-RC1
- AC2: 'Reply on RC1', Albrecht Weerts, 08 Apr 2024
  
  We like to thank the reviewer for his/her feedback. We think the review will help to sharpen our manuscript
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC2
RC2:
'Comment on hess-2023-274', Anonymous Referee #2, 29 Dec 2023

This manuscript investigates the potential changes in flash flood frequency and magnitude to be expected in the Alpine regions following the ongoing climate change. The study setup is based on convection-permitting model simulations that are fed to a hydrological model.
The ability of this chain to reproduce flash flooding in the past is evaluated and the performance of this setup is compared with the one obtained by feeding global-scale renalyses to the hydrological model.
The study investigates one of the key questions concerning precipitation-driven hazards under climate change and adopts state-of-the-art modelling elements that I deem adequate to the task. The authors put a huge amount of work in the study, and the manuscript is well written and clear.
At the same time, the simulation setup comes short at addressing some aspects that I believe are important in shaping the results. My concerns mainly revolve around the two points below. Addressing them may require abundant work, including additional analyses. I am not sure this can be done within a major revision.
(1) The assumption of trading space for time in the case of flash floods needs to be better motivated. Trading space for time is generally used for processes that are not scale-dependent. Flash floods are, as they can be generated by diverse meteorological forcing, in relation to the basin area and characteristics. This means that we can have nearby basins that respond to different meteorological forcing with different spatial and temporal scales. These different forcing can be represented with different accuracy by climate models.

This aspect is neglected in the flash flood validation. For example, it is possible to have a peak in a larger nearby catchment 2 days apart from the original event due to completely different meteorological forcing (an incoming large-scale front as opposed to isolated convection, for instance).

On this aspect, how is 3 days chosen (see line 198)? It appears a long time to me, unless there are known uncertainties in the timing of the used databases.
(2) Several processes in flash flood generation are non-linear. Therefore, quantitative values of many variables are critical for flash flood response. I feel that the model setup neglects the biases in the CPM simulations. Quantitative biases are expected to be enhanced by the non-linear hydrological processes typical of flash flood generation. While it is true that there a lot of attention in the validation of the modelling chain (section 2.6), the performance of said chain are not very high (e.g. see figure 3).

- How much can we trust the modelling chain quantitatively? Is the relatively low performance related to deficiencies in the CPM simulations or in the hydrological model? Perhaps some of these answers can be replied to by exploring the simulated precipitation fields and comparing them to the triggering ones. For example, in the case of the database from Amponsah et al, radar estimates for all the events are provided.

- The text in lines 301-303 attempts an explanation on why there is no adjustment but, in light of the non-linear responses mentioned above, I believe prudence should be put in saying that comparing model with model decreases the importance of the biases.

- It is not clear to me how is the hydrological model calibrated. Is it calibrated based on the CPM simulations? ERA5? ERA-Interim? I believe the calibration strategy should be better described as it also plays an important role in shaping the results.
Minor comments:
- Fig. 6,7: the reported variance is very large. It could be related to regional scale variability and scale-dependence (area) in the climate change response. The organisation into large-scale fluvial catchments may mask these aspects. Do you see any spatial pattern? Is it possible to organise it somehow into regions with homogeneous response? And/or by basin area?
- In Fig 6, I’d suggest using a log transformation on the y axis

Citation: https://doi.org/10.5194/hess-2023-274-RC2
- AC1: 'Reply on RC2', Albrecht Weerts, 08 Apr 2024
  
  Publisher’s note: the content of this comment was removed on 17 April 2024 since the comment was posted by mistake.
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC1
- AC4: 'Reply on RC2', Albrecht Weerts, 08 Apr 2024
  
  We like to thank the reviewer. We think the review will help to sharpen our manuscript.
  Unfortunately we uploaded the wrong reply. This one contains the correct reply.
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC4
RC3:
'Comment on hess-2023-274', Martina Kauzlaric, 11 Jan 2024
General comments and recommendation
The manuscript submitted by Zander et al. presents the results of a study aiming at quantifying changes in flash flood frequency in the Alpine domain related to “worst case” climate change (RCP 8.5). The authors make use of the results of recent convention-permitting regional climate models’ (CP-RCMs) hourly simulations with a 2.2 km spatial resolution and feed these to a distributed hydrological model set up for the five largest basins outspringing in the central European Alps, with a cell size of roughly 1 km².
The hydrological model is not calibrated, if I understand it correctly, but parameters are rather inferred from (pedo)transfer functions. The modeling chain is first validated and evaluated using common/already existing global reanalysis products (daily ERA5 and hourly CP-RCM-downscaled ERA-Interim) as inputs, and continuous discharge measurements together with peak specific discharge events extracted from two flash flood data bases (HANZE and EuroMedeFF) as observations. In a second step changes in flash flood frequency are investigated, inferred from the simulations performed using the raw CP-RCMs data (i.e. no downscaling resp. correction is undertaken for CP-RCMs data, but a bare remapping to the resolution of the hydrological model is performed).
The manuscript as a whole is well structured, the methods are generally understandable or mostly supported by relevant sources, the discussion is good, but quite (too?) short. I think some parts of the latter rather belong to the methods part (e.g. the choice of not bias correcting the data). In general, some important assumptions/concepts/choices/etc.., as well as the novelty of the study could (should?) be better elaborated and highlighted.
I am not a native speaker, and as such not the best judge, but I dare to say the paper is mostly written correctly, and it mostly reads fine, just sometimes is not consistent in the terminology (e.g. Current climate and Historical climate referring to the same dataset).
The quality of the figures featured in the manuscript could be definitely improved, while the relative captions accompanying these are clear and provide good descriptions.
Despite the main idea of the study is in fact really interesting and to my knowledge has not yet been addressed with such a framework, I have several concerns about some assumptions made and the implementation of this analysis:
The hypothesis of trading space for time: flash floods are a specific subcategory of floods. A flash flood is usually a quite localized (10-100 km²), short-lasting event (min to hr), in order to have a more widespread effect and have a regional character it needs some “ingredients”, such as special weather patterns (for the European Alps probably an atmospheric river or slow-moving storms) and/or special preconditions, such as saturated soils. The time window of potential occurrence and the time steps used in this study for validation/evaluation are way too large for this kind of floods (3 days and hourly data aggregated to mean daily discharge, respectively), and don’t really allow to make any conclusion on the goodness and the suitability of the modelling framework applied here.

The definition of flash flood used in fact completely lacks the inclusion of the time component, it is highly questionable if this assumption can be justified.

The lack of the observed specific discharges and of the location of the flash floods for the nine recorded flash floods used in the validation: In Table 2 only modelled peak discharges are reported, and often the whole subbasin is provided as the region being affected, even though more precise indications are available in the data bases used. E.g. the events on the 6^th of June and on the 8^th of September in the Rhone subbasin: both events took place in France (6^th of June in Isère and 8^th of September in Ardèche and further downstream) considerably distant from the outlet of the Rhone subbasin actually modelled (Rhone down to Geneva), for which only two gauging stations are used. The swiss part of the Rhone in 2002 experienced only locally some relevant floods in November, as well as the Ticino, to my knowledge. Or can the authors prove better?

Discharge data: the discharge data are spatially extremely biased. Why is that? There are considerably more stations available also for the Rhine, the Rhone and the Po. In the manuscript there is no description of how the choice of the stations was made, and also a more informative overview of the stations used is missing (e.g. with the size of the subcatchment, the length of observations, the river and the corresponding basin). This could be easily provided also in Appendix or as Support Material.

The F1 and Peirce skill scores are actually quite poor. The reason and the dynamics behind these results are not really further investigated, so that is difficult to understand and trust, what the model is doing.

10 years reference time frame for inferring changes in frequency: is really a decade a period long enough to make any inference on changes in frequency of flashfloods as compared to the current period? (this is part of the original hypothesis of trading space for time) The authors cite studies on extreme value statics from short time series of rainfall, but floods and in particular flash floods occur significantly less frequently than heavy precipitation (see e.g. Modrick and Georgakakos, 2015). One might also argue, that the results might be strongly influenced by the chosen decade, depending if it falls within a flood-rich or a flood-poor period (see also Lun at al.2020, Fischer et al. 2023). A discussion and a climatological embedding are missing.

No bias correction/ downscaling of climate data: while it might be true that applying a correction is not always necessary when switching meteorological product (see Reed et al. 2007 and Alfieri&Thielen, 2015), it is however difficult to judge in this study if this is a valid choice, because there is no evaluation from the climatological-hydrological point of view of the “historical” period (current period). Objectively, we only know from Ban et al.2021, Fig.7 that in France and Italy summer bias in frequency of hourly precipitation is larger than -25%, while in Switzerland this bias is between -5 and -25%, and in fall is between +25-40% when downscaling ERA-Interim with UM.

Because of these considerations, I think the manuscript requires substantial revisions and additional work, before it can be considered for publication.
Please find my specific and technical comments here following.
Specific comments
Abstract:
Please be consistent in the terminology, either you always write Flash Floods with two capital letters (as the first two words in the abstract), or you stick to flash floods (then make the F in floods small)

Introduction:
P2-L27: Kotlarski et al.(2023) is not listed in the References..

I miss an embedding in what is currently into place on a European level for simulating and detecting flash floods. Being one of the deadliest flood (hazard) type, quite some things and studies were performed on the continental scale in the last decade. E.g. https://www.efas.eu/en/flash-flood-indicators or Ritter al al. 2021?

Data and Method:
P3-L87: what do you mean with ridge height?

Being a swiss hydrologist and seeing your study area is basically covering the whole of hydrological Switzerland, and this makes probably also most of your study area (?), another useful database with peak specific discharge to be used when analyzing your flash floods on a regional scale could be found here https://opendata.swiss/de/dataset/grosse-hochwasserabfluesse-in-der-schweiz (however most of the explanations are available only in German, I am afraid). As I said already before, we miss an overview of the (numeric) areas of the basins and also of their sub-basins, it is important to know the spatial scale, in particular because you make use of area as a threshold and also of peak specific discharge as the main output variable.

1: why do you use OpenStreetMap to produce this figure?( and also Fig. 2 later on) It would be easier (and nicer?) to produce these figures yourself.

P5-L119: Chan et al. (2020) is also missing in the References.

P5-L121: is a one in a year event an extreme event? I would reformulate this.

P5-L127-129: I am not sure I understand why this is so not “unambiguous” (it has been done many times, simply assuming clear-sky conditions and solving a nice equation using latitude, day of the year –for estimating solar declination- and hour angle?). If I understand it correctly you assume there is radiation also in the night? I believe Osnabrugge et al. 2019 used hourly downward radiation.

P6-L143: wouldn’t it be easier if you would simply say you used a priori PTFs and upscaling rules as Imhoff et al.(2020)? Most hydrologists would then know you used something similar to an MPR.

P6-L147: Myeni et al.2015 is only pertinent for MODIS data, I think, for CORINE data please cite properly Copernicus.

P6-L160: why do you use ERA5 as forcing to perform the sensitivity analysis, and in general for model validation? I don’t understand why you take a daily dataset - kind of setting it as a benchmark- while all the rest is performed on an hourly basis. What is the rationale? Why didn’t you use ERA5 Land for example, which is available at fine spatial and temporal scales (hourly, about 9km), more similar to the meteorological products you used in the rest of the study?

P6-161-162: why do you choose specifically these two stations specifically to conduct a sensitivity analysis on a factor that is applied to the whole modelling domain? The Thur river is nested in the catchment of the Rhine at Basel, which is per se not bad, but 1) why don’t you perform this same analysis at other stations in different climatic regions? 2) The Rhine at Basel might be quite strongly influenced (and the sensitivity reduced?) by the several large regulated lakes upstream, did you consider it when making your choice?

P6-L162: based on which criteria was a factor 100 defined as satisfactory?

Please avoid saying downscaled while you actually simply remapped data, it is misleading (e.g. P7 L167)

P7-L176: why do you only use KGE for validation? I don’t think it is appropriate to use only KGE as a goodness-of-fit measure for validating a modelling framework devoted to flash floods (in particular if you only use mean daily discharges!)

What is the spin-up period of the climatological models? And of the hydrological model? The spin-up period is not mentioned anywhere, I believe?

P7-L210: you sometimes switch “Historical” and “Current”, please be consistent throughout the manuscript (I would suggest to always use “Current” and remove the reference “Historical”).

Results:
It would be very beneficial if you would provide more detailed information on the selected floods (i.e. extend Table 2), and show some hydrographs of how this flash floods were picked (or not) by the model.

I would remove the grey background in all figures.

P9-L221-223: In Figure 3 you show one random year of simulation for two catchments, this is not really telling us too much about the model performance and ability to pick up the annual cycle or snowmelt overall. The Thur is a subalpine (transition) catchment (so snowmelt is relatively limited), and the Ticino in Bellinzona is quite strongly influenced by hydropower upstream (and the largest summer events are all missed?).

7: maybe making additionally a plot per basin with subcatchment size on the x-axes would help to see if the very large values come e.g. from very small catchments? It is difficult to disentangle the actual meaning and consequences of the results presented in Fig.7. And also, if the results would look the same if e.g. precipitation intensity would be aggregated and plotted for the same basin and subcatchment subdivision.

Discussion:
Did you consider that flash floods might be also shifting in time/season in the future?(see e.g. Blöschl et al. 2017 and Tarasova et al. 2023)

In general the discussion could be expanded..

Technical corrections
Throughout the text and in some caption you use “&”, please avoid doing this and use the word “and” instead, which is more appropriate for a paper.

Please see the attached document for more technical corrections.

References
Alfieri, L., Thielen, J. (2015): A European precipitation index for extreme rain-storm and flash flood early warning; Meteorol. Appl., 22(1), 3–13, https://doi.org/10.1002/met.1328
Blöschl, G. et al.,(2017): Changing climate shifts timing of European floods; Science 357,588-590(2017);DOI:10.1126/science.aan2506
Fischer, S., Lun, D., Schumann A.H., Blöschl, G. (2023): Detecting flood-type-specific flood-rich and flood-poor periods in peaks-over-threshold series with application to Bavaria (Germany); Stochastic Environmental Research and Risk Assessment 37, 1395 – 1413, https://doi.org/10.1007/s00477-022-02350-8
Lun, D., Fischer, S., Viglione, A., & Blöschl, G. (2020): Detecting Flood‐Rich and Flood‐Poor Periods in Annual Peak Discharges Across Europe; Water Resources Research, 56(7). https://doi.org/10.1029/2019wr026575
Modrick, T.M., Georgakakos, K. P. (2015): The character and causes of flash flood occurrence changes in mountainous small basins of Southern California under projected climatic change; Journal of Hydrology: Regional Studies, Vol. 3, 312-336, ISSN 2214-5818, https://doi.org/10.1016/j.ejrh.2015.02.003
Reed, S., Schaake, J., Zhang, Z. (2007): A distributed hydrologic model and threshold frequency-based method for flash flood forecasting at ungauged locations. J. Hydrol. 337: 402–420. https://doi.org/10.1016/j.jhydrol.2007.02.015
Ritter, J., Berenguer, M., Park, S., Sempere-Torres, D. (2021): Real-time assessment of flash flood impacts at pan-European scale: The ReAFFINE method; Journal of Hydrology, Vol. 603, Part C, 127022, ISSN 0022-1694, https://doi.org/10.1016/j.jhydrol.2021.127022
Tarasova, L., Lun, D., Merz, R., Blöschl, G., Basso, S., BErtola, M, Miniusii, A., Rakovec, O., Samaniego, L., Thober, S., Kumar, R. (2023): Shifts in flood generation processes exacerbate regional flood anomalies in Europe. Commun Earth Environ 4, 49; https://doi.org/10.1038/s43247-023-00714-8
Citation: https://doi.org/10.5194/hess-2023-274-RC3
- AC3: 'Reply on RC3', Albrecht Weerts, 08 Apr 2024
  
  We like to thank the reviewer for his/her comments. We think the comments will help to improve the manuscript
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC3

Status: closed

RC1:
'Comment on hess-2023-274', Anonymous Referee #1, 22 Dec 2023
This study investigates the potential changes in flash floods over the Alps due to climate change. The aim of this study was to evaluate the use of a convection-permitting regional climate model in combination with a distributed hydrological model to assess future changes in the frequency and magnitude of flash floods over the Alps. The UM convection-permitting model was used to project future climate at high spatial and temporal resolution. These projections were used as inputs to the distributed hydrological model wflow_sbm. The ability of the hydrological model to simulate historical flash floods was first assessed using reanalysis data. The modelling framework was then used to investigate future changes in flash floods.
Although the topic of the paper is very interesting and has not yet been addressed in the literature, I do not believe that the methodology is adequate to address the research question formulated in the introduction. The modelling framework and data set used in this study are not robust enough to support the conclusions.
Major comments:
The ability of the hydrological model to reproduce historical flash floods in the study area is very low. Apart from the fact that the methodology for assessing this is questionable (see comments below), the results show that the model is not able to detect floods based on the threshold approach for about half of the stations studied (Fig. 5). The hydrographs in Figure 3 also show the limited performance in simulating most floods (and not just flash floods). If the modelling framework does not capture the dominant processes triggering floods in this region in the historical period, it is very unlikely that it will do so under changing conditions. There could be several reasons for this (e.g. uncertainty in the input data). It could also be argued that a better modelling framework has not yet been developed. However, since the current modelling capabilities (at least for the modelling framework of this study) are not robust and reliable enough in this context, I do not think it should be used to study future changes in flash floods.

The definition of a flash flood lacks a very important feature: the temporal dynamics of the storm and/or flood event. To determine what is a flash flood, the authors use the specific peak discharge and the size of the upstream catchment. This definition was adapted from Amponsah et al. (2018) by removing the storm duration as a selection criterion (not mentioned in the manuscript). This means that any discharge exceeding the specified threshold and occurring in a catchment of less than 3000 km² is considered a flash flood, regardless of whether it was triggered by a uniform precipitation event lasting a few days or by a very intense and highly spatially variable convective storm event. Floods associated with slow catchment dynamics can theoretically be considered as flash floods in this modelling framework.

The validation/evaluation methodology is not suitable for assessing the ability of the modelling framework to project future changes in flash floods.
The aggregation to the daily time step for model validation does not allow to evaluate the ability of the model to simulate flash floods, because the hydrological processes involved affect the flow at sub-daily time steps.

Basing the performance assessment on peak flows and the KGE calculated over the entire time series is not sufficient to assess whether the modelling framework was able to capture the dominant flood generation processes.
As the catchments do not necessarily only experience flash floods, the number of points related to these floods used to calculate an error criterion is small compared to the rest of the time series. Therefore, a high KGE value in this case does not mean that the model performs well in simulating flash floods. In addition, the KGE values are quite low for many stations (KGE< 0.6 for about ½ of the stations, see Table 3, Crochemore et al. 2015).

Peak flow analyses rely on single points, which can be highly uncertain.

As part of the validation/evaluation analysis is based on nine historical floods, it would have been possible to plot the flood hydrographs and compare the simulations with the observations (if available).

The paper lacks a comprehensive description of the streamflow dataset, in particular why this dataset is suitable for answering the research question. It is not clear why only 130 gauged stations were used in this study and how they were selected.
Were the stations selected because their catchments experience flash floods? Were they chosen because of data availability, low human influence...?

What are the hydrological processes affecting river flows in these catchments?

Are all the catchments and sub-catchments prone to flash floods and how often compared to the other floods?

Why were stations with different temporal resolutions chosen?

A large part of the study area is not covered by gauged stations.

No bias correction was applied to the climate projections. This point is discussed in Section 4, the main arguments being that bias correction can distort the change signal and that the observational data available over the study region do not have the resolution of the CPM. However, most hydrological impact studies apply some form of bias correction, as biases are usually quite large at the hydrological scale (e.g. Teutschbein and Seibert, 2012). It is not clear how the choice not to apply a bias correction affects the results of this study, as there is no assessment of the climate variables for the historical period compared to the observations. Are the biases large enough to significantly change peak discharge simulations?

An analysis of changes in flood drivers might have been expected. Other studies reporting on the potential drying of the region are mentioned (l276 to 280), but the states and fluxes of the hydrological model could have been used for a more in-depth analysis.

As I understand it, one of the objectives of this study was to assess the potential added value of using a CPM in combination with a distributed hydrological model to assess future changes in flash floods over the Alps. In order to assess the added value of such a framework, this study should have included a benchmark, such as the projections of a regional climate model at 0.11° resolution.

Other comments
Some of the figures and tables do not highlight the results well enough (suggestions for improving the figures below).
2: add the delineation of the study region.

3: one hydrograph per line; extend in width; eventually show a smaller period.

5: increase resolution; change “f1” and “peirce_ss” to “F1” and “Peirce”. Add number of stations in the legend. Add number of stations with skill score = 0.

2: add a column with observed peak specific discharge.

6: apply a transformation on the values of the x-axis to improve visualisation.

L11-12: “and to determine a suitable threshold definition for flash flooding.” I did not understand why the definition of a threshold for flash floods is one of the objectives and why it is based on modelling results and not on observations.

L148-156: why is glacier modelling mentioned in these lines? The authors show that glaciers do not have a significant influence on the occurrence of flash floods, but do not conclude on the implications for their modelling framework.

L 198: the regional approach to flash flood validation should be better explained, e.g. with a simple figure showing how the threshold approach is used to determine whether the model detects a flash flood or not.

L205: “rare extreme events”. If the flash floods considered in this study are rare events, the limited duration of the CPM projections (10 years) may not allow such events to be studied. I would have expected a small analysis of the flood events that occurred in the historical period for the 130 stations.

L 268: “5.6m3s−1km−2 compared to 4.49m3s−1km−2”. The difference between the two values is most likely within the uncertainty range of the hydrological model (to be confirmed with the observed peak values to be added to Table 2).

L276 to 280 could be moved to the discussion section.

References:
Louise Crochemore, Charles Perrin, Vazken Andréassian, Uwe Ehret, Simon P. Seibert, Salvatore Grimaldi, Hoshin Gupta & Jean-Emmanuel Paturel (2015) Comparing expert judgement and numerical criteria for hydrograph evaluation, Hydrological Sciences Journal, 60:3, 402-423, DOI: 10.1080/02626667.2014.903331
Teutschbein, C., & Seibert, J. (2012) Bias correction of regional climate model simulations for hydrological climate-change impact studies: Review and evaluation of different methods. Journal of hydrology, 456, 12-29, DOI: 10.1016/j.jhydrol.2012.05.052
Citation: https://doi.org/10.5194/hess-2023-274-RC1
- AC2: 'Reply on RC1', Albrecht Weerts, 08 Apr 2024
  
  We like to thank the reviewer for his/her feedback. We think the review will help to sharpen our manuscript
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC2
RC2:
'Comment on hess-2023-274', Anonymous Referee #2, 29 Dec 2023

This manuscript investigates the potential changes in flash flood frequency and magnitude to be expected in the Alpine regions following the ongoing climate change. The study setup is based on convection-permitting model simulations that are fed to a hydrological model.
The ability of this chain to reproduce flash flooding in the past is evaluated and the performance of this setup is compared with the one obtained by feeding global-scale renalyses to the hydrological model.
The study investigates one of the key questions concerning precipitation-driven hazards under climate change and adopts state-of-the-art modelling elements that I deem adequate to the task. The authors put a huge amount of work in the study, and the manuscript is well written and clear.
At the same time, the simulation setup comes short at addressing some aspects that I believe are important in shaping the results. My concerns mainly revolve around the two points below. Addressing them may require abundant work, including additional analyses. I am not sure this can be done within a major revision.
(1) The assumption of trading space for time in the case of flash floods needs to be better motivated. Trading space for time is generally used for processes that are not scale-dependent. Flash floods are, as they can be generated by diverse meteorological forcing, in relation to the basin area and characteristics. This means that we can have nearby basins that respond to different meteorological forcing with different spatial and temporal scales. These different forcing can be represented with different accuracy by climate models.

This aspect is neglected in the flash flood validation. For example, it is possible to have a peak in a larger nearby catchment 2 days apart from the original event due to completely different meteorological forcing (an incoming large-scale front as opposed to isolated convection, for instance).

On this aspect, how is 3 days chosen (see line 198)? It appears a long time to me, unless there are known uncertainties in the timing of the used databases.
(2) Several processes in flash flood generation are non-linear. Therefore, quantitative values of many variables are critical for flash flood response. I feel that the model setup neglects the biases in the CPM simulations. Quantitative biases are expected to be enhanced by the non-linear hydrological processes typical of flash flood generation. While it is true that there a lot of attention in the validation of the modelling chain (section 2.6), the performance of said chain are not very high (e.g. see figure 3).

- How much can we trust the modelling chain quantitatively? Is the relatively low performance related to deficiencies in the CPM simulations or in the hydrological model? Perhaps some of these answers can be replied to by exploring the simulated precipitation fields and comparing them to the triggering ones. For example, in the case of the database from Amponsah et al, radar estimates for all the events are provided.

- The text in lines 301-303 attempts an explanation on why there is no adjustment but, in light of the non-linear responses mentioned above, I believe prudence should be put in saying that comparing model with model decreases the importance of the biases.

- It is not clear to me how is the hydrological model calibrated. Is it calibrated based on the CPM simulations? ERA5? ERA-Interim? I believe the calibration strategy should be better described as it also plays an important role in shaping the results.
Minor comments:
- Fig. 6,7: the reported variance is very large. It could be related to regional scale variability and scale-dependence (area) in the climate change response. The organisation into large-scale fluvial catchments may mask these aspects. Do you see any spatial pattern? Is it possible to organise it somehow into regions with homogeneous response? And/or by basin area?
- In Fig 6, I’d suggest using a log transformation on the y axis

Citation: https://doi.org/10.5194/hess-2023-274-RC2
- AC1: 'Reply on RC2', Albrecht Weerts, 08 Apr 2024
  
  Publisher’s note: the content of this comment was removed on 17 April 2024 since the comment was posted by mistake.
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC1
- AC4: 'Reply on RC2', Albrecht Weerts, 08 Apr 2024
  
  We like to thank the reviewer. We think the review will help to sharpen our manuscript.
  Unfortunately we uploaded the wrong reply. This one contains the correct reply.
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC4
RC3:
'Comment on hess-2023-274', Martina Kauzlaric, 11 Jan 2024
General comments and recommendation
The manuscript submitted by Zander et al. presents the results of a study aiming at quantifying changes in flash flood frequency in the Alpine domain related to “worst case” climate change (RCP 8.5). The authors make use of the results of recent convention-permitting regional climate models’ (CP-RCMs) hourly simulations with a 2.2 km spatial resolution and feed these to a distributed hydrological model set up for the five largest basins outspringing in the central European Alps, with a cell size of roughly 1 km².
The hydrological model is not calibrated, if I understand it correctly, but parameters are rather inferred from (pedo)transfer functions. The modeling chain is first validated and evaluated using common/already existing global reanalysis products (daily ERA5 and hourly CP-RCM-downscaled ERA-Interim) as inputs, and continuous discharge measurements together with peak specific discharge events extracted from two flash flood data bases (HANZE and EuroMedeFF) as observations. In a second step changes in flash flood frequency are investigated, inferred from the simulations performed using the raw CP-RCMs data (i.e. no downscaling resp. correction is undertaken for CP-RCMs data, but a bare remapping to the resolution of the hydrological model is performed).
The manuscript as a whole is well structured, the methods are generally understandable or mostly supported by relevant sources, the discussion is good, but quite (too?) short. I think some parts of the latter rather belong to the methods part (e.g. the choice of not bias correcting the data). In general, some important assumptions/concepts/choices/etc.., as well as the novelty of the study could (should?) be better elaborated and highlighted.
I am not a native speaker, and as such not the best judge, but I dare to say the paper is mostly written correctly, and it mostly reads fine, just sometimes is not consistent in the terminology (e.g. Current climate and Historical climate referring to the same dataset).
The quality of the figures featured in the manuscript could be definitely improved, while the relative captions accompanying these are clear and provide good descriptions.
Despite the main idea of the study is in fact really interesting and to my knowledge has not yet been addressed with such a framework, I have several concerns about some assumptions made and the implementation of this analysis:
The hypothesis of trading space for time: flash floods are a specific subcategory of floods. A flash flood is usually a quite localized (10-100 km²), short-lasting event (min to hr), in order to have a more widespread effect and have a regional character it needs some “ingredients”, such as special weather patterns (for the European Alps probably an atmospheric river or slow-moving storms) and/or special preconditions, such as saturated soils. The time window of potential occurrence and the time steps used in this study for validation/evaluation are way too large for this kind of floods (3 days and hourly data aggregated to mean daily discharge, respectively), and don’t really allow to make any conclusion on the goodness and the suitability of the modelling framework applied here.

The definition of flash flood used in fact completely lacks the inclusion of the time component, it is highly questionable if this assumption can be justified.

The lack of the observed specific discharges and of the location of the flash floods for the nine recorded flash floods used in the validation: In Table 2 only modelled peak discharges are reported, and often the whole subbasin is provided as the region being affected, even though more precise indications are available in the data bases used. E.g. the events on the 6^th of June and on the 8^th of September in the Rhone subbasin: both events took place in France (6^th of June in Isère and 8^th of September in Ardèche and further downstream) considerably distant from the outlet of the Rhone subbasin actually modelled (Rhone down to Geneva), for which only two gauging stations are used. The swiss part of the Rhone in 2002 experienced only locally some relevant floods in November, as well as the Ticino, to my knowledge. Or can the authors prove better?

Discharge data: the discharge data are spatially extremely biased. Why is that? There are considerably more stations available also for the Rhine, the Rhone and the Po. In the manuscript there is no description of how the choice of the stations was made, and also a more informative overview of the stations used is missing (e.g. with the size of the subcatchment, the length of observations, the river and the corresponding basin). This could be easily provided also in Appendix or as Support Material.

The F1 and Peirce skill scores are actually quite poor. The reason and the dynamics behind these results are not really further investigated, so that is difficult to understand and trust, what the model is doing.

10 years reference time frame for inferring changes in frequency: is really a decade a period long enough to make any inference on changes in frequency of flashfloods as compared to the current period? (this is part of the original hypothesis of trading space for time) The authors cite studies on extreme value statics from short time series of rainfall, but floods and in particular flash floods occur significantly less frequently than heavy precipitation (see e.g. Modrick and Georgakakos, 2015). One might also argue, that the results might be strongly influenced by the chosen decade, depending if it falls within a flood-rich or a flood-poor period (see also Lun at al.2020, Fischer et al. 2023). A discussion and a climatological embedding are missing.

No bias correction/ downscaling of climate data: while it might be true that applying a correction is not always necessary when switching meteorological product (see Reed et al. 2007 and Alfieri&Thielen, 2015), it is however difficult to judge in this study if this is a valid choice, because there is no evaluation from the climatological-hydrological point of view of the “historical” period (current period). Objectively, we only know from Ban et al.2021, Fig.7 that in France and Italy summer bias in frequency of hourly precipitation is larger than -25%, while in Switzerland this bias is between -5 and -25%, and in fall is between +25-40% when downscaling ERA-Interim with UM.

Because of these considerations, I think the manuscript requires substantial revisions and additional work, before it can be considered for publication.
Please find my specific and technical comments here following.
Specific comments
Abstract:
Please be consistent in the terminology, either you always write Flash Floods with two capital letters (as the first two words in the abstract), or you stick to flash floods (then make the F in floods small)

Introduction:
P2-L27: Kotlarski et al.(2023) is not listed in the References..

I miss an embedding in what is currently into place on a European level for simulating and detecting flash floods. Being one of the deadliest flood (hazard) type, quite some things and studies were performed on the continental scale in the last decade. E.g. https://www.efas.eu/en/flash-flood-indicators or Ritter al al. 2021?

Data and Method:
P3-L87: what do you mean with ridge height?

Being a swiss hydrologist and seeing your study area is basically covering the whole of hydrological Switzerland, and this makes probably also most of your study area (?), another useful database with peak specific discharge to be used when analyzing your flash floods on a regional scale could be found here https://opendata.swiss/de/dataset/grosse-hochwasserabfluesse-in-der-schweiz (however most of the explanations are available only in German, I am afraid). As I said already before, we miss an overview of the (numeric) areas of the basins and also of their sub-basins, it is important to know the spatial scale, in particular because you make use of area as a threshold and also of peak specific discharge as the main output variable.

1: why do you use OpenStreetMap to produce this figure?( and also Fig. 2 later on) It would be easier (and nicer?) to produce these figures yourself.

P5-L119: Chan et al. (2020) is also missing in the References.

P5-L121: is a one in a year event an extreme event? I would reformulate this.

P5-L127-129: I am not sure I understand why this is so not “unambiguous” (it has been done many times, simply assuming clear-sky conditions and solving a nice equation using latitude, day of the year –for estimating solar declination- and hour angle?). If I understand it correctly you assume there is radiation also in the night? I believe Osnabrugge et al. 2019 used hourly downward radiation.

P6-L143: wouldn’t it be easier if you would simply say you used a priori PTFs and upscaling rules as Imhoff et al.(2020)? Most hydrologists would then know you used something similar to an MPR.

P6-L147: Myeni et al.2015 is only pertinent for MODIS data, I think, for CORINE data please cite properly Copernicus.

P6-L160: why do you use ERA5 as forcing to perform the sensitivity analysis, and in general for model validation? I don’t understand why you take a daily dataset - kind of setting it as a benchmark- while all the rest is performed on an hourly basis. What is the rationale? Why didn’t you use ERA5 Land for example, which is available at fine spatial and temporal scales (hourly, about 9km), more similar to the meteorological products you used in the rest of the study?

P6-161-162: why do you choose specifically these two stations specifically to conduct a sensitivity analysis on a factor that is applied to the whole modelling domain? The Thur river is nested in the catchment of the Rhine at Basel, which is per se not bad, but 1) why don’t you perform this same analysis at other stations in different climatic regions? 2) The Rhine at Basel might be quite strongly influenced (and the sensitivity reduced?) by the several large regulated lakes upstream, did you consider it when making your choice?

P6-L162: based on which criteria was a factor 100 defined as satisfactory?

Please avoid saying downscaled while you actually simply remapped data, it is misleading (e.g. P7 L167)

P7-L176: why do you only use KGE for validation? I don’t think it is appropriate to use only KGE as a goodness-of-fit measure for validating a modelling framework devoted to flash floods (in particular if you only use mean daily discharges!)

What is the spin-up period of the climatological models? And of the hydrological model? The spin-up period is not mentioned anywhere, I believe?

P7-L210: you sometimes switch “Historical” and “Current”, please be consistent throughout the manuscript (I would suggest to always use “Current” and remove the reference “Historical”).

Results:
It would be very beneficial if you would provide more detailed information on the selected floods (i.e. extend Table 2), and show some hydrographs of how this flash floods were picked (or not) by the model.

I would remove the grey background in all figures.

P9-L221-223: In Figure 3 you show one random year of simulation for two catchments, this is not really telling us too much about the model performance and ability to pick up the annual cycle or snowmelt overall. The Thur is a subalpine (transition) catchment (so snowmelt is relatively limited), and the Ticino in Bellinzona is quite strongly influenced by hydropower upstream (and the largest summer events are all missed?).

7: maybe making additionally a plot per basin with subcatchment size on the x-axes would help to see if the very large values come e.g. from very small catchments? It is difficult to disentangle the actual meaning and consequences of the results presented in Fig.7. And also, if the results would look the same if e.g. precipitation intensity would be aggregated and plotted for the same basin and subcatchment subdivision.

Discussion:
Did you consider that flash floods might be also shifting in time/season in the future?(see e.g. Blöschl et al. 2017 and Tarasova et al. 2023)

In general the discussion could be expanded..

Technical corrections
Throughout the text and in some caption you use “&”, please avoid doing this and use the word “and” instead, which is more appropriate for a paper.

Please see the attached document for more technical corrections.

References
Alfieri, L., Thielen, J. (2015): A European precipitation index for extreme rain-storm and flash flood early warning; Meteorol. Appl., 22(1), 3–13, https://doi.org/10.1002/met.1328
Blöschl, G. et al.,(2017): Changing climate shifts timing of European floods; Science 357,588-590(2017);DOI:10.1126/science.aan2506
Fischer, S., Lun, D., Schumann A.H., Blöschl, G. (2023): Detecting flood-type-specific flood-rich and flood-poor periods in peaks-over-threshold series with application to Bavaria (Germany); Stochastic Environmental Research and Risk Assessment 37, 1395 – 1413, https://doi.org/10.1007/s00477-022-02350-8
Lun, D., Fischer, S., Viglione, A., & Blöschl, G. (2020): Detecting Flood‐Rich and Flood‐Poor Periods in Annual Peak Discharges Across Europe; Water Resources Research, 56(7). https://doi.org/10.1029/2019wr026575
Modrick, T.M., Georgakakos, K. P. (2015): The character and causes of flash flood occurrence changes in mountainous small basins of Southern California under projected climatic change; Journal of Hydrology: Regional Studies, Vol. 3, 312-336, ISSN 2214-5818, https://doi.org/10.1016/j.ejrh.2015.02.003
Reed, S., Schaake, J., Zhang, Z. (2007): A distributed hydrologic model and threshold frequency-based method for flash flood forecasting at ungauged locations. J. Hydrol. 337: 402–420. https://doi.org/10.1016/j.jhydrol.2007.02.015
Ritter, J., Berenguer, M., Park, S., Sempere-Torres, D. (2021): Real-time assessment of flash flood impacts at pan-European scale: The ReAFFINE method; Journal of Hydrology, Vol. 603, Part C, 127022, ISSN 0022-1694, https://doi.org/10.1016/j.jhydrol.2021.127022
Tarasova, L., Lun, D., Merz, R., Blöschl, G., Basso, S., BErtola, M, Miniusii, A., Rakovec, O., Samaniego, L., Thober, S., Kumar, R. (2023): Shifts in flood generation processes exacerbate regional flood anomalies in Europe. Commun Earth Environ 4, 49; https://doi.org/10.1038/s43247-023-00714-8
Citation: https://doi.org/10.5194/hess-2023-274-RC3
- AC3: 'Reply on RC3', Albrecht Weerts, 08 Apr 2024
  
  We like to thank the reviewer for his/her comments. We think the comments will help to improve the manuscript
  
  Citation: https://doi.org/10.5194/hess-2023-274-AC3

Marjanne J. Zander, Pety J. Viguurs, Frederiek C. Sperna Weiland, and Albrecht H. Weerts

Viewed

Total article views: 515 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
347	137	31	515	12	16

HTML: 347
PDF: 137
XML: 31
Total: 515
BibTeX: 12
EndNote: 16

Views and downloads (calculated since 05 Dec 2023)

Month	HTML	PDF	XML	Total
Dec 2023	220	68	9	297
Jan 2024	43	20	3	66
Feb 2024	15	13	2	30
Mar 2024	24	18	4	46
Apr 2024	45	18	13	76

Cumulative views and downloads (calculated since 05 Dec 2023)

Month	HTML	PDF	XML	Total
Dec 2023	220	68	9	297
Jan 2024	43	20	3	66
Feb 2024	15	13	2	30
Mar 2024	24	18	4	46
Apr 2024	45	18	13	76

Viewed (geographical distribution)

Total article views: 505 (including HTML, PDF, and XML) Thereof 505 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 26 Apr 2024

Short summary

Flash floods are damaging natural hazard which often occur in the European Alps. High resolution climate model output is combined with high resolution distributed hydrological models to model changes in flash flood frequency and intensity. Results show a similar flash flood frequency for autumn in the future, but a decrease in summer. However, the future discharge simulations indicate an increase in the flash flood severity in both summer and autumn leading to more severe flash flood impacts.


Total:	0
HTML:	0
PDF:	0
XML:	0